Predicting Whether Surgery Will Relieve Seizures
A model applied to neuroimaging data revealed biomarkers that may predict outcomes in patients with epilepsy.
Many patients with temporal lobe epilepsy stand to benefit from surgery, which provides better relief from seizures than medications alone. Yet 30-50 percent of these patients continue to experience seizures even after surgery, underscoring the need for better methods to predict individual treatment responses and to improve clinical decision-making. In a study published Dec. 5 in Frontiers in Neuroscience, researchers tackled this challenge by developing a statistical model to identify whole-brain biomarkers from preoperative positron emission tomography imaging, known as PET. In theory, the biomarkers will predict post-surgical seizure recurrence after resection of the anterior temporal lobe.
“The question of why some patients do well after surgery and others do not is something which is not yet understood,” said first author Sharon Chiang of Rice University and Baylor College of Medicine in Houston, Texas. “If we could better predict which patients have the greatest likelihood of benefit, those with a low probability of benefiting from surgery could potentially be considered for other noninvasive alternatives to resective surgery.”
Traditionally, the prediction of post-surgical outcomes using PET has focused on specific regions within the temporal lobe. But increasing evidence suggests that temporal lobe epilepsy is a network disorder that affects brain regions beyond the temporal lobe. This suggests that whole-brain statistical approaches applied to neuroimaging data may reveal features that are reliably associated with individual clinical outcomes.
Motivated by this prospect, Chiang and her collaborators applied their modeling strategy to preoperative PET data collected from 19 adult patients with drug-resistant mesial temporal lobe epilepsy who underwent anterior temporal lobe resection at the University of California, Los Angeles Seizure Disorder Center. The analysis also included preoperative functional MRI data collected from a separate set of 32 temporal lobe epilepsy patients at the same center.
“Statistical methods which allow neurologists and neurosurgeons to use information about brain networks from fMRI to identify subtle signal abnormalities on positron emission tomography are needed,” Chiang said. “We provide a novel approach to integrating the information in these two imaging modalities to predict who will or will not benefit from surgery.”
This approach revealed that the 19 patients who underwent surgery could be divided into two subgroups associated with different clinical outcomes, with one subgroup having a 5.8 times greater chance of postoperative seizure recurrence after anterior temporal lobe resection. These two subgroups were characterized by different levels of glucose metabolism in various brain regions within and beyond the temporal lobe.
In addition to identifying subgroups of subjects and selecting imaging biomarkers, the modeling approach provides a probabilistic estimate of an individual patient's risk of postoperative seizure recurrence, with a prediction accuracy of 84 percent. But according to senior author Marina Vannucci of Rice University, “more extensive testing on larger datasets and from multi-institutional patient populations is needed before real-time clinical usage.”
Donald Berry, a biostatistician at the University of Texas M.D. Anderson Cancer Center in Houston, said that “the investigators’ model is not ready for prime time.” In addition to being validated in independent patient cohorts, the findings must also “be compared with other prognostic measures, including the radiologist’s prediction of the results of surgery based on his or her subjective assessment of the PET scans,” he explained.
One strength of the modeling approach is that it “builds a framework for how to combine all the parts of the pre-surgical evaluation into a single, objective evaluation for the benefits and risks of surgery,” said Wesley Kerr, a mathematician and medical intern at Eisenhower Medical Center in Rancho Mirage, California, who managed the PET records for the study. But he agreed that more patients are needed. “In order to really train the algorithm and understand its broader impact on clinical care, the authors need at least 100 subjects to train and 50 subjects to validate their predictions prospectively,” he said. “Previous work has shown that the good performances we see on small sample sizes often are not reproduced on larger sample sizes.”